Maternal Work Hours and Adolescents’ School Outcomes Among Low-Income Families in Four Urban Counties

Abstract

We examine how changes in maternal work hours affect adolescent children’s school participation and performance outcomes using data from interviews in 1998 and 2001 with approximately 1,700 women who, in May 1995, were welfare-reliant, single mothers of adolescents living in neighborhoods of concentrated poverty in Cuyahoga (Cleveland), Los Angeles, Miami-Dade, and Philadelphia counties. Analyses control for a broad array of mothers’ characteristics, including their psychological and physical health, experiences with domestic violence and substance abuse, as well as unobserved time-invariant characteristics. In fixed-effects models, we find unfavorable effects of increased maternal work hours on three of six outcomes: skipping school, performing above average, and parental contact about behavior problems. Adolescent-aged sons seem to be particularly sensitive to changes in mothers’ hours of work.

Low-income youth face an array of barriers to academic attendance, achievement, and high school completion, which in turn affects their opportunities for ongoing secondary education and, ultimately, their labor market success (Duncan and Brooks-Gunn 1997; MacLeod 1995). Indeed, increased human capital among youth is critical to future economic security and reducing their dependence on cash assistance, both of which are key policy goals related to at-risk youth (Card 1999). Unlike youth from more advantaged families, low-income youth are more likely to be raised by single mothers (Edin and Kefalas 2005), to live in resource-poor and unsafe neighborhoods (Brooks-Gunn, Duncan, and Aber 1997), attend low-quality schools (Kozol 1991), to be members of economically unstable families (Scott et al. 2004), and to face discrimination and differential treatment in a variety of contexts (Fine 2005; Lareau 2003). Welfare-reliant mothers of these youth are increasingly under pressure to leave the welfare rolls and patch together financial resources through their own wage labor and that of other family members (Blank and Haskins 2001; London, Scott, and Hunter 2002). Many struggle to balance the new opportunities, demands, and limitations associated with Temporary Assistance for Needy Families (TANF) and the exigencies of the low-wage labor market with their children’s needs for attention, care, and supervision (London et al. 2004; Scott, London, and Hurst 2005; Weigt 2006).

How youth in these circumstances fare while mothers change the amount of time they spend in paid employment—and struggle to achieve or maintain independence from cash assistance—is relatively understudied (Duncan and Chase-Lansdale 2001). Most research on this question has focused on young children because they are viewed as especially susceptible developmentally to the potential costs or benefits of maternal employment (Chase-Lansdale et al. 2003; Morris et al. 2001; Waldfogel, Han, and Brooks-Gunn 2002). In this paper, we examine how change in maternal work hours affects adolescent children’s school outcomes by using data from interviews in 1998 and 2001 with approximately 1,700 women who, in May 1995, were 18- to 45-year-old, welfare-reliant, single mothers living in the most-disadvantaged neighborhoods in four urban counties: Cuyahoga (Cleveland), Los Angeles, Miami-Dade, and Philadelphia.

This study has several features that contribute to the extant literature. Methodologically, we take advantage of a rarely available, rich set of covariates to control for maternal physical and psychological health, domestic violence, substance abuse, and several aspects of socioeconomic well-being and family structure. With two waves of data, we also employ a fixed-effects technique that controls for time-invariant, unobserved characteristics. Substantively, the effect of average hours of maternal employment on adolescents’ school participation and performance outcomes is assessed linearly as well as trichotomously, depicting part-time work (1 to 30 hours) separately from full-time work (31 hours or more per week), with no hours of paid employment as the reference category. We additionally assess whether the effects of maternal employment differ for adolescent-aged sons and daughters (Eccles 1999) and conduct a number of supplemental analyses to enhance our understanding of the effects we observe.

BACKGROUND AND CONCEPTUAL MOTIVATION

We draw on theories spanning a variety of social science disciplines to inform hypotheses about how changes in parents’ labor force participation can affect children’s development generally, and school outcomes specifically. From the perspective of economic theory, parental employment can affect children’s development by influencing the amount and distribution of parental resources and time investments (Becker 1981; Becker and Tomes 1979, 1986; Ruhm 2006). All else being equal, if increased labor force participation yields increased income, then one would expect greater parental investment in their children’s human capital. At the same time, working more hours constrains the time available for children (Kurz 2002; London et al. 2004; Milkie et al. 2004); however, parents may arrange their nonwork hours so that quality time with their children is not reduced (Bianchi 2000; Bryant and Zick 1996; Chase-Lansdale et al. 2003; Zick, Bryant, and Osterbacka 2001). Economic theory also predicts that older youth assess the likely returns to continued education versus immediate employment—an assessment that might be altered by the parent’s ability to support the family without financial contributions from the child.

Sociological and psychological theories point to parenting as one mechanism by which maternal employment, and its subsequent influences on parental emotional well-being, can affect interactions between parents and their children (Chase-Lansdale and Pittman 2002; Parcel and Menaghan 1994). If mothers find jobs that are rewarding and challenging, then they might also experience decreases in psychological distress; decreased maternal depression could in turn translate into increased warmth and responsiveness in parenting, which is known to be beneficial for children’s development (Moore and Driscoll 1997; Zaslow, McGroder, and Moore 2000). Alternatively, pressure to find work, securing a low-quality job (e.g., with inflexible schedules, tedious tasks, and low wages), or working more hours in an unsatisfying job can increase stress, reduce emotional well-being, and increase harsh and less responsive or less consistent parenting (Hofferth et al. 2000; McCloyd 1990). Working parents may simply have fewer psychosocial resources available to devote to good parenting even if they are spending the same amount of time with their children as they did when they were not working (Baumrind 1991; Menaghan and Parcel 1995). Increases in nonstandard work hours might affect parents’ knowledge and monitoring of their adolescent’s lives, which in turn is linked to better school outcomes (Baker and Stevenson 1986; Baumrind 1989; Kerr and Stattin 2000; Sampson and Laub 1994). Developmental psychologists and sociologists also emphasize the importance of socialization in children’s attainment. Mothers who work model normative behaviors relevant to labor force attachment for their children and thereby serve as role models (London et al. 2004).

Developmental psychology additionally pays attention to the timing of changes in parent-child relationships and family life across a child’s developmental lifespan. Spending less time at home because of work may lead parents to expect adolescents to take on new “adult” tasks, which could lead to increased responsibility (Dodson and Dickert 2004; Hsueh and Gennetian 2006) and better behavior. For some adolescents, this could also lead to resentment, acting out, failure to complete unsupervised tasks (such as homework), and resistance to any kind of control imposed by an adult (Burton, Brooks, and Clark 2002).

Changes in maternal work could affect sons’ and daughters’ development and school participation and performance differently (Eccles 1999). When hours of employment increase, mothers may rely more on their adolescent-aged daughters than on their sons to help with household chores and responsibilities (Crouter et al. 2001). Parents may also invest differently in the future education of their sons and daughters (Butcher and Case 1994).

Most of the available research on the effect of maternal employment on adolescents examines two-parent or middle- to higher-income families and generally finds few or slightly favorable effects on outcomes such as adolescent achievement and teen childbearing (e.g., see Aughinbaugh and Gittleman 2004; Bogenschneider and Steinberg 1994; Lopoo 2004). Consistent with these general findings, the few studies that have explicitly focused on low-income, single mother samples—particularly during the 1990s, a period during which there were dramatic declines in welfare caseloads and increases in the labor force participation of single mothers—have also found neutral to favorable associations between maternal employment and adolescent self-esteem and academic achievement (Allesandri 1992; Kalil, Dunifon, and Danziger 2001; McLoyd et al. 1994; Ruhm 2006). These relationships appear to be quite robust in national, regional, and urban area samples, and are observed regardless of whether prior or current welfare receipt of the parent is considered (e.g., see Chase-Lansdale et al. 2003).

In contrast, research based on a synthesis of experimental studies of welfare and work programs has found that welfare and work policies, per se, produce unfavorable effects on a range of schooling outcomes among adolescents, particularly those who have younger siblings at home (Gennetian et al. 2004). A somewhat complementary finding also emerged in work using data from the Panel Study of Income Dynamics (PSID), which showed that stringent welfare policies, measured at the state level over time, increased the rate of dropping out of school (Hao, Astone, and Cherlin 2004). Additionally, a recent study that used a more nuanced characterization of maternal work found that adolescents with single mothers who are in “bad” jobs (i.e., full-time jobs with low pay and no health insurance) experienced more grade repetition (Kalil and Ziol-Guest 2005). As this brief review of prior work suggests, there is some uncertainty in the literature regarding how increased maternal employment can affect low-income adolescents’ school participation and performance.

SAMPLE, DATA, AND MEASURES

Data for this study are from the Project on Devolution and Urban Change (henceforth, Urban Change), a longitudinal study conducted in Cuyahoga (Cleveland), Los Angeles, Miami-Dade, and Philadelphia counties in the early years of the implementation of the welfare reforms brought about by the passage in 1996 of the Personal Responsibility and Work Opportunity Reconciliation Act (PRWORA). These four urban counties accounted for approximately 14% of the entire U.S. welfare caseload as of 1999 (Allen and Kirby 2000).

The Urban Change survey involved women who, in May 1995, were single mothers between the ages of 18 and 45 years, were receiving cash assistance (Aid to Families with Dependent Children) and/or Food Stamps, and were living in census tracts where either the poverty rate exceeded 30% or the rate of welfare receipt exceeded 20%.1 From administrative records comprising the entire caseload who met these criteria, approximately 1,000 women were randomly sampled in each site. First-round interviews were completed between March 1998 and February 1999, with a response rate of 79%.2 An analysis of response bias indicated that whites who were sampled were significantly less likely than African Americans to have been interviewed in 1998, while women with more children in their households were significantly more likely to have been interviewed; however, in both instances, the differences were small (Polit et al. 2001).

Second-round interviews were conducted in the spring of 2001. Of the 3,960 women who were surveyed in 1998, 3,260 were reinterviewed in 2001, for an overall retention rate of 82%. Results of an analysis of attrition bias show that respondents who completed the 2001 interview did not statistically differ from nonresponders across a variety of demographic characteristics in Cuyahoga County (Brock et al. 2002). In Miami, however, whites, Hispanics, persons under the age of 25 years, and widows were less likely to complete the 2001 interview (Brock et al. 2004), while in Philadelphia, there was differential response by marital status (Michalopoulos et al. 2003). In Philadelphia, separated women were most likely to complete the 2001 interview, followed by single women, and finally by married women. In Los Angeles, the full set of demographic predictors did not significantly differentiate responders from nonresponders, although Latinas were significantly more likely than African Americans to respond to the 2001 interview, and women with a GED but no college were significantly less likely to respond than women with a high school diploma but no college (Polit et al. 2005).

The women who participated in the Urban Change survey provided detailed information about a wide range of topics, including their experiences with welfare reform and use of support and safety net services; employment histories, wages, hours, and income; family configurations, living arrangements, child care, and parenting; perceptions of their neighborhoods; experiences of material hardships; health status; and experiences of domestic violence. In addition to providing selected information about the well-being of every biological or adopted child in the household, all of the female survey respondents provided in-depth information about two pre-designated focal children: a Focal Child A (age 2–6 years in 1998) and a Focal Child B (age 12–18 years in 1998).

As part of a larger project, we converted the original, mother-level Urban Change data set for each year into a multiyear, child-level data set. Because detailed information was obtained for a range of school participation and performance outcomes and there is a relative dearth of rich data on low-income youth and their parents, this study focuses on those outcomes for adolescents who are Focal B children. For the analytical sample, we select only those Focal B children who were less than 19 years old and living in their mother’s household at both interviews, whose mothers were interviewed in both 1998 and 2001, and who had information reported at both interviews for at least one of the schooling outcomes. In 1998, the ages of the adolescents ranged from 12 to 16 years, with the majority being age 14 or less (mean age = 13.4, SD = 1.2). In 2001, the ages of adolescents ranged from 14 to 18 years, with the majority being age 17 or less (mean = 15.9, SD = 1.1). Satisfaction of these criteria leaves a maximum potential sample of 1,698 child-year observations: 958 in 1998 and 740 in 2001.3 As detailed in the Appendix, the Focal B adolescents who were excluded from the analytic sample because they did not meet the inclusion criteria appear to be worse off across numerous adolescent and maternal indicators than those who were included in our analyses.4

This study examines a set of school-related outcomes that we group broadly into participation and performance outcomes. The participation outcomes include maternal reports regarding (1) whether the child had skipped school or cut classes without permission in the prior 12 months, (2) the number of days of school missed in the prior four weeks, and (3) the number of days late to school in the prior four weeks.5 The maternally reported performance outcomes include an overall current school performance variable scaled from “1: Not Well at All” to “5: Very Well,” with 3 set to “Average,” and an indicator for whether the mother had been contacted about attendance/behavior/academic problems in the prior 12 months. From the overall school performance variable, we created two indicator variables to capture thresholds of above- and below-average performance. We code above-average performance equal to 1 if the mother reported a rating of “4: Well” or “5: Very Well,” and 0 otherwise. We set below-average performance equal to 1 if the mother reported a rating of “2: Below Average” or “1: Not Well At All,” and 0 otherwise. All of the outcomes are measured for only those adolescents who were attending school at both survey waves. Three of these outcome measures are similarly collected by the U.S. Department of Education for the Digest of Education Statistics, allowing us to draw a comparison with a national sample of 10th graders in 2002 (U.S. Department of Education 2005). Roughly 10% of the adolescents in our sample missed more than 6 days of school in one month, compared with 17.2% of 10th graders who report missing more than 6 days of school over a much longer time period, that is, the first half of the current school year. Nearly 31% of parents of 10th graders nationally were contacted about problems their child was having with school, compared with 41% in our full analysis sample.

Urban Change collected information on maternal employment for up to four jobs. For each of these positions, respondents were asked, “Including overtime, how many hours per week (do/did) you usually work on this job?” We created the maternal work hours measure by summing the responses to this question for all current jobs. Among the mothers who were working at the time of the 2001 survey, the majority, 87.4%, had one job, while 10.3% had two jobs. In the models below, we examine both a linear and a categorical specification (i.e., a specification with indicators for working 30 or fewer hours and working more then 30 hours, with no work as the reference category). We focused on a 30-hour threshold to address questions of practical policy significance, since state TANF programs’ weekly work or work-related activities requirements are often at 30 or more hours (Parrott et al. 2006).6

The empirical models include the child’s sex, race/ethnicity, and age, as well as the mother’s age, educational attainment, marital status, cohabitation status, and place of birth (i.e., within the United States or not). We also include several indicators of maternal health and well-being, including her SF-12 physical health component score (Ware, Kosinski, and Keller 1996), her score on the Center for Epidemiological Studies-Depression (CES-D) scale (Radloff 1977), an indicator for whether a health condition limits her ability to work, indicators for whether the mother has been physically or emotionally abused in the past year,7 and an indicator for whether she reported using a hard drug in the past month.8 We also include measures for the presence of another adult in the household, the presence of a child other than the respondent’s own or adopted children, and the number of children in the household. Finally, we include site indicators, as well as interactions between site and the Hispanic and U.S.-born indicators, respectively. See Table 1 for a list of covariates with values for the study sample.

Table 1.

Schooling Outcomes of Low-Income Youth, Descriptive Statistics of the Sample^a

Variable	Wave 1 1998		Wave 2 2001		Full Analysis Sample
	Mean	SD	Mean	SD	Mean	SD
Maternal Employment in Survey Month
Average weekly hours	17.7	21.0	23.6	22.0	20.3	21.6
% worked ≤ 30 hours	13.4		14.1		13.7
% worked > 30 hours	34.2		45.9		39.3
Child Outcomes
Number of days late for school in past four weeks	1.4	2.6	1.9	3.1	1.6	2.8
Number of days missed school in past four weeks	2.0	2.9	2.0	3.0	2.0	2.9
% skipped school/class without permission	14.0		23.3		18.0
% performed above average in school	49.9		55.4		52.3
% performed below average in school	19.7		16.3		18.2
% parent contacted about behavior problems in school	43.5		36.6		40.5
Characteristics of Children
% male	49.8		50.0		49.9
Age	13.4	1.2	15.9	1.1	14.5	1.7
Characteristics of Mothers
Age	36.1	5.0	38.6	5.0	37.2	5.2
% no high school diploma or equivalent	48.3		43.4		46.2
% currently married	8.3		14.9		11.2
% currently cohabiting	9.6		12.1		10.7
CES-D scoreb	17.9	11.3	16.2	10.5	17.2	11.0
% with health limitation	25.5		24.3		25.0
SF-12 physical component scorec	46.2	9.9	46.0	10.2	46.1	10.0
% physically abused (past year)	8.3		4.5		6.6
% emotionally abused (past year)	40.3		26.1		34.1
% used hard drugs (past 30 days)	2.4		2.0		2.2
% born in the United States	78.8		78.1		78.5
% black	68.9		68.8		68.9
% Hispanic	25.0		25.6		25.2
% white	4.5		3.9		4.3
% other	1.6		1.6		1.6
Characteristics of Household
% with “other” adult in householdd	24.0		26.4		25.0
% with “other” children in householde	10.0		14.6		12.0
Total number of children in household	3.1	1.6	3.1	1.6	3.1	1.6
Total Sample Size (OLS)	958		740		1,698
Total Number of Children	958		740		958

^aThe sample is children aged 12–18 at Wave 1 and under age 19 at Wave 2.

^bCES-D = Center for Epidemiologic Studies-Depression scale; scores can range in value from 0 to 60. A score above 23 is considered an indicator of a high risk of depression.

^cSF-12 is a health survey with 12 questions designed to measure mental and physical health. Physical component scores have been standardized in a national sample to a mean of 50 and a standard deviation of 10.

^dOther than mother, spouse, or cohabitating partner.

^eOther than respondent’s own or adopted children.

To maximize our sample size, we use ordinary least squares (OLS) regression to impute values for the covariate values that are missing (other than maternal work hours). In the imputation process, we include all of the other covariates as well as the year from which the data were drawn. We include a set of indicator variables (one for each covariate with over 10 values assigned) in all models, which are set to 1 if the value for a given variable was imputed. The findings are nearly identical with and without these controls for missing data on the covariates.

ANALYTIC APPROACH

Our aim in this work is to capitalize on the longitudinal variation in maternal work hours and adolescent outcomes such that we can estimate the contemporaneous effects of maternal work hours on adolescent school participation and performance outcomes. We begin with a reduced-form specification—using OLS or logit techniques as appropriate—modeling outcome y for adolescent i in year t, controlling for a matrix of time-invariant (Z) and time-varying factors (X):

$$y_{it}\hspace{0.17em}=\hspace{0.17em}\alpha +{\beta }_{1}EM{P}_{it}+{X^{\prime }}_{it}{\beta }_2+{Z^{\prime }}_i{\beta }_3+{\varepsilon }_{it},$$ (1)

where EMP represents a measure of maternal work hours.

Eq. (1) yields unbiased and consistent estimates of β₁ if cov(ɛ, EMP) = 0. This condition is unlikely because a variety of unobserved factors are likely to create an association between maternal work hours and adolescent outcomes. For example, increases in maternal work hours might be facilitated when adolescents are highly competent, responsible, or mature. In contrast, persistent emotional/social problem behavior might reduce maternal work hours. In the absence of experimental data, perhaps the best econometric approach to estimate unbiased coefficients for maternal employment is an instrumental variables model. Unfortunately, we were unable to identify a variable that is correlated to the endogenous variable (maternal work hours) that was plausibly independent of the error term. Given the close relationship between human capital accumulation and labor force participation, most factors that would induce change in mother’s work would also probably influence an adolescent’s school participation and performance.

To control for time-invariant, unobserved characteristics, we have chosen to use a fixed-effects model for the continuous outcomes and Chamberlain’s (1980) fixed-effects logit model for the binary outcomes.9 More specifically, these models remove the time-invariant fixed effect from the error term (η_i):

$$y_{it}=\mu +{\gamma }_{1}EM{P}_{it}+{X^{\prime }}_{it}{\gamma }_2+{Z^{\prime }}_i{\gamma }_3+{\eta }_i+{\varepsilon }_{it},$$ (2)

Operationally, these fixed-effects models use variation in the outcome and maternal employment within a given family over time to identify the maternal employment coefficient. Because these models use variation within a mother-adolescent dyad, characteristics that are time invariant cannot be the explanation for changes in the outcome; thus, the coefficient for maternal employment should not suffer from bias due to unobserved time-invariant characteristics. Below, we further discuss the variation over time in maternal employment and the strength of our model in detecting this variation.

Although they are an improvement over OLS or logit models, fixed-effects models have limitations. As mentioned earlier, one needs variation in the outcome to identify the coefficient for maternal work hours in all fixed-effects models. In fact, the likelihood function for the fixed-effects logit requires variation in the outcome across waves; individuals who do not have a different value for the outcome in the two observed years are omitted from the identification process. If individuals with no variation in the outcome over time are different from those who do have variation over time, then our results are not generalizable. In addition, even though we are fortunate to have access to a broad array of covariates at both time points—including maternal physical and psychological health, measures of emotional and physical abuse, and drug use—other unmeasured omitted factors that vary over time and that are correlated with both maternal work hours and the outcome might continue to produce bias in the maternal work hours coefficients.

FINDINGS

For ease of exposition, we group the educational outcomes into two categories: school participation and school performance. Table 2 reports results using a linear specification for average hours of maternal work on these outcomes and presents estimates obtained from OLS or logit models and from fixed-effects models.

Table 2.

Estimates of the Effects of Average Maternal Work Hours on Adolescent School Participation and Performance Outcomes

Outcome	OLS	Fixed Effects	Logit	Fixed-Effects Logit
Number of Days Late for School	0.002	−0.002
SE	(0.004)	(0.006)
N	1,423	1,030
Number of Days Missed School	−0.001	0.008
SE	(0.004)	(0.006)
N	1,452	1,076
Skipped School/Class Without Permission			0.003	0.024**
SE			(0.004)	(0.009)
Marginal effects			[0.000]	[0.004]
N			1,668	316
Performed Above Average in School			−0.002	−0.015*
SE			(0.003)	(0.006)
Marginal effects			[−0.001]	[−0.004]
N			1,691	476
Performed Below Average in School			0.000	−0.000
SE			(0.003)	(0.007)
Marginal effects			[0.000]	[−0.000]
N			1,691	310
Parent Contacted About Behavior Problems			0.004	0.011*
SE			(0.003)	(0.006)
Marginal effects			[0.001]	[0.003]
N			1,695	530

Notes: Logit and fixed-effects logit models are used to predict the dichotomous outcomes. Standard errors are in parentheses; the cluster option was used in all of these models to adjust standard errors for the lack of independence between observations (multiple observations per child). Marginal effects are in brackets; the formula is (∂P / ∂EMP) = P × (1 – P) × γ₁. The following controls are included: child’s sex and age; mother’s age; educational attainment; SF-12 physical health component score; CES-D score; the number of children in the household; and indicators for marital and cohabitation status, race/ethnicity, whether or not a health condition limits the mother’s ability to work, whether or not the mother was born in the United States, whether or not the mother has recently been physically abused, emotionally abused, or has used hard drugs, the presence of another adult in the household, the presence of a child other than the mother’s own/adopted child, the respondent’s location, a set of interactions between each site location and the Hispanic indicator, and a set of interactions between the site locations and an indicator for born in the United States. Also included are indicators flagging missing values for each of the covariates.

^*p < .05;

^**p < .01.

In Table 2, the initial OLS or logit models for all of the participation outcomes show trivial and statistically nonsignificant effects for maternal work hours. The fixed-effects models, however, illustrate the importance of omitted variable bias. While the maternal employment coefficient remains statistically unrelated to the number of days missed and number of days tardy, the coefficient estimate is positive and statistically significant for skipping school. In addition to the coefficient estimates for the logit models, we also report marginal effects (in brackets), or the change in the probability of the outcome given a one-unit change in maternal work hours for the logit models. The marginal effect for maternal work hours in the skipped school model is 0.4 percentage points.10

Initial results for the school performance outcomes follow a similar pattern: we find no statistically significant relationship between maternal employment and school performance outcomes in the logit models. However, we do find statistically significant coefficients in two of the fixed-effects models. Although maternal employment does not appear to influence the probability that a child performs poorly in school, it is related to the probability that the adolescent will perform well in school with a marginal effect of −0.4 percentage points. We also find a positive, statistically significant relationship between maternal employment and parents being contacted for behavior problems. A one-hour increase in maternal work hours is associated with a 0.3-percentage-point increase in the probability that a parent was contacted by the school.

By using a linear maternal work hours variable in Table 2, we implicitly assume that each additional hour of work will produce the same change in the outcome. In Table 3, we report results from fixed-effects models with the categorical specification of maternal work hours. Changing the measure of maternal work hours does not alter the null findings for the “days missed” and “days tardy” outcomes. We do find that maternal employment is statistically significantly related to skipping school regardless of the mother’s work status. Compared with the children of nonworking mothers, the children of mothers who work between 1 and 30 hours are 20 percentage points more likely to skip school (i.e., more than twice as likely as the proportion in the sample). Similarly, compared with the adolescent children of nonworking mothers, the children of mothers who work 31 or more hours are 16 percentage points more likely to skip school. The effect of working between 1 and 30 hours does not statistically differ from the effect of working 31 or more hours. For the performance measures—more specifically, for the “above-average performance” outcome—the categorical specification of employment shows that the negative relationship for maternal employment surfaces in a statistically discernible way only when mothers work more than 30 hours. Compared with the children of nonworking mothers, the children of mothers who work more than 30 hours per week are 17 percentage points less likely to perform above average in school. We also find a significant positive coefficient for mothers who work more than 30 hours in the “parent contacted about problem behavior” model, and this coefficient statistically differs from the coefficient on mothers who worked 30 hours or less. The marginal effect for these families is 14 percentage points, or an increase of 35% over the sample proportion.

Table 3.

Fixed-Effects Estimates of the Effects of Full- and Part-Time Work on Adolescent Schooling Outcomes

Variable	Number of Days Late for School	Number of Days Missed School	Skipped School/Class Without Permission	Performed Above Average in School	Performed Below Average in School	Parent Contacted About Behavior Problems in School
Worked 1 to 30 hours	−0.129	−0.056	1.369*	−0.334	0.717	−0.347
SE	(0.347)	(0.338)	(0.639)	(0.339)	(0.459)	(0.324)
Marginal effects			[0.202]	[−0.083]	[0.107]	[−0.084]
Worked > 30 hours	−0.121	0.518	1.082*	−0.666*	0.020	0.589*
SE	(0.299)	(0.322)	(0.505)	(0.270)	(0.343)	(0.295)
Marginal effects			[0.160]	[−0.166]	[0.003]	[0.142]

Notes: Fixed-effects logit models were used to predict the dichotomous outcomes. Standard errors are in parentheses; the cluster option was used in all of these models to adjust standard errors for the lack of independence between observations (multiple observations per child). Marginal effects are in brackets; the formula is (∂P / ∂EMP) = P × (1 – P) × γ₁. The following controls are included: child’s sex and age; mother’s age; educational attainment; SF-12 physical health component score; CES-D score; the number of children in the household; and indicators for marital and cohabitation status, race/ethnicity, whether or not a health condition limits the mother’s ability to work, whether or not the mother was born in the United States, whether or not the mother has recently been physically abused, emotionally abused, or has used hard drugs, the presence of another adult in the household, the presence of a child other than the mother’s own/adopted child, the respondent’s location, a set of interactions between each site location and the Hispanic indicator, and a set of interactions between the site locations and an indicator for born in the United States. Also included are indicators flagging missing values for each of the covariates

^*p < .05

Do the effects of maternal employment on any of these school participation and performance outcomes differ for sons compared with daughters? Table 4 shows results from fixed-effects models that include interactions between a categorical specification for work hours and an indicator variable equal to 1 if the adolescent is male. Prior findings showed that maternal employment influences the probability that adolescents skipped school and that the mother was contacted about problem behavior. The results reported in Table 4 suggest that these relationships do not differ by the child’s sex. However, we do find evidence that the relationship between maternal employment and the number of days tardy and performing above average is different for sons and daughters. Compared with the adolescent children of nonworking mothers, the sons of mothers who work 1 to 30 hours were late about 1.2 days more in the prior four weeks than the daughters of mothers who work the same number of hours. The gender difference, including the magnitude of the difference, is the same for the adolescents of mothers who work 31 or more hours per week. Compared with the children of nonworking mothers, sons of mothers who work 1 to 30 hours are 28 percentage points less likely to perform above average than daughters of mothers who work comparable hours.

Table 4.

Fixed-Effects Estimates of the Effects of Work Hours Interacted With Adolescent Sex on Schooling Outcomes

Variable	Number of Days Late for School	Number of Days Missed School	Skipped School/Class Without Permission	Performed Above Average in School	Performed Below Average in School	Parent Contacted About Behavior Problems in School
Worked 1 to 30 hours	−0.691	−0.416	0.494	0.290	−0.469	−0.933†
SE	(0.496)	(0.494)	(1.445)	(0.548)	(0.916)	(0.518)
Marginal effects			[0.073]	[0.072]	[−0.07]	[−0.225]
Worked > 30 hours	−0.698	0.453	1.527*	−0.259	−0.397	0.207
SE	(0.425)	(0.458)	(0.678)	(0.356)	(0.545)	(0.365)
Marginal effects			[0.225]	[−0.065]	[−0.060]	[0.05]
Worked 1 to 30 hours × Male	1.177†	0.723	1.347	−1.119†	1.609	1.012
SE	(0.696)	(0.669)	(1.781)	(0.677)	(1.121)	(0.676)
Marginal effects			[0.199]	[−0.279]	[0.240]	[0.244]
Worked > 30 hours × Male	1.245*	0.122	−0.926	−0.850	0.645	0.832
SE	(0.591)	(0.609)	(1.049)	(0.534)	(0.702)	(0.586)
Marginal effects			[−0.137]	[−0.212]	[0.096]	[0.200]

Notes: Fixed-effects logit models were used to predict the dichotomous outcomes. Standard errors are in parentheses; the cluster option was used in all of these models to adjust standard errors for the lack of independence between observations (multiple observations per child). Marginal effects are in brackets; the formula is (∂P / ∂EMP) = P × (1 – P) × γ₁. The following controls are included: child’s sex and age; mother’s age; educational attainment; SF-12 physical health component score; CES-D score; the number of children in the household; and indicators for marital and cohabitation status, race/ethnicity, whether or not a health condition limits the mother’s ability to work, whether or not the mother was born in the United States, whether or not the mother has recently been physically abused, emotionally abused, or has used hard drugs, the presence of another adult in the household, the presence of a child other than the mother’s own/adopted child, the respondent’s location, a set of interactions between each site location and the Hispanic indicator, and a set of interactions between the site locations and an indicator for born in the United States. Also included are indicators flagging missing values for each of the covariates.

^†p < .10;

^*p < .05

SUPPLEMENTAL ANALYSES

Employment Transitions

In our preferred models, the fixed-effects specification, we identify the relationship between maternal work hours and the adolescent schooling outcomes using adolescent-specific variation over time. Since intraindividual variation is crucial to the model, we were concerned about the generalizability of the model if all of the variation in maternal employment was confined to a particular portion of the distribution of work hours (e.g., if the only changes observed were among mothers increasing work hours from 25 to 31 hours). In order to assess this concern in our analyses, we constructed a “transition matrix” for the maternal employment variables in our panel data. Because it has the largest sample size, we broke down the analytical sample from the “contacted parents for problem behavior” model into the proportion of the adolescents with mothers who worked 0, between 1 and 30, and more than 30 hours in 1998 (see Table 5). In 1998, 52% of the mothers did not work, 13% worked between 1 and 30 hours, and 35% worked more than 30 hours per week on average. Of the mothers who were not working in 1998, 57% were still not working in 2001. Of the 43% who were working in 2001, 15% were working between 1 and 30 hours, and 28% were working more than 30 hours. Of the mothers who were working 1 to 30 hours in 1998, 24% were no longer working in 2001, 25% were still working 1 to 30 hours, and 51% were working more than 30 hours. Finally, of the women working more than 30 hours in 1998, 72% were still working more than 30 hours in 2001, 20% were no longer working, and roughly 9% were working 1 to 30 hours. These descriptive results show that employment transitions among the women in the sample occurred throughout the distribution of maternal employment hours with variation coming from both increases and decreases in maternal employment over time.

Table 5.

Changes in Maternal Work Hours Across Survey Waves

	Wave 2 (2001)			Employment Status Wave 1 (Row Total)	N
	0 Hours	1–30 Hours	31+ Hours
Wave 1 (1998)
0 Hours	0.57	0.15	0.28	0.52	380
1–30 Hours	0.24	0.25	0.51	0.13	95
31+ Hours	0.20	0.09	0.72	0.35	256
Total				1.00	731

Sample Composition or Fixed Effect?

The fixed-effects models show some statistically significant relationships compared with the OLS and logit models. We follow Lopoo (2005) and conduct a supplemental analysis to ascertain the extent to which the results reflected the removal of bias via the fixed effect or the composition of the analytic subsample. As explained earlier, the fixed-effects models identify the maternal employment coefficient using only the adolescent observations with a change in their schooling outcome over time. In Table 6, we report coefficients for the nonlinear work hours specification estimated using a logit model and a fixed-effects model on both the full sample and the initial analytic subsample used for the fixed-effects models. We confine our results to the three outcomes for which we find statistically significant coefficients. Our results indicate that both the removal of the fixed effect and the changes in sample composition contribute to the findings.

Table 6.

Comparing Logit and Fixed-Effects Logit Models Using Full Sample and Subsamples With Outcome Variation for Selected Schooling Outcomes

Variable	Skipped School/Class Without Permission			Performed Above Average in School			Parent Contacted About Behavior Problems
	Logit		Fixed-Effects Logit	Logit		Fixed-Effects Logit	Logit		Fixed-Effects Logit
	Full Sample	Fixed-Effects Logit Subsample	Fixed-Effects Logit Sample	Full Subsample	Fixed-Effects Logit Subsample	Fixed-Effects Logit Sample	Full Subsample	Fixed-Effects Logit Subsample	Fixed-Effects Logit Subsample
Worked 1 to 30 hours	0.058	0.741†	1.369*	−0.178	−0.326	−0.334	0.129	−0.176	−0.347
SE	(0.212)	(0.420)	(0.639)	(0.165)	(0.297)	(0.339)	(0.167)	(0.299)	(0.324)
Marginal effects	[0.009]	[0.109]	[0.202]	[−0.044]	[−0.080]	[−0.083]	[0.031]	[−0.042]	[−0.084]
Worked > 30 hours	0.074	0.570†	1.082*	−0.078	−0.530*	−0.666*	0.128	0.385*	0.589*
SE	(0.212)	(0.330)	(0.505)	(0.127)	(0.210)	(0.270)	(0.126)	(0.196)	(0.295)
Marginal effects	[0.011]	[0.084]	[0.160]	[−0.019]	[−0.132]	[−0.166]	[0.031]	[0.093]	[0.142]
N	1,668	316	316	1,691	476	476	1,695	530	530

Notes: Logit and fixed-effects logit models are used to predict the dichotomous outcomes. Standard errors are in parentheses; the cluster option was used in all of these models to adjust standard errors for the lack of independence between observations (multiple observations per child). Marginal effects are in brackets; the formula is (∂P / ∂EMP) = P × (1 – P) × γ₁. The following controls are included: child’s sex and age; mother’s age; educational attainment; SF-12 physical health component score; CES-D score; the number of children in the household; and indicators for marital and cohabitation status, race/ethnicity, whether or not a health condition limits the mother’s ability to work, whether or not the mother was born in the United States, whether or not the mother has recently been physically abused, emotionally abused, or has used hard drugs, the presence of another adult in the household, the presence of a child other than the mother’s own/adopted child, the respondent’s location, a set of interactions between each site location and the Hispanic indicator, and a set of interactions between the site locations and an indicator for born in the United States. Also included are indicators flagging missing values for each of the covariates.

^†p < .10;

^*p < .05

For example, in the first set of columns, we report results from three different models of the “skipped school/class without permission” outcome. In the first column, we report results using the full Urban Change sample (N = 1,668) and a logit model with the same specification as reported in Table 2. Neither coefficient is statistically significant. Next, we report results using a logit model with the same specification using a sample that has variation in the outcome (i.e., the analytic subsample used for the fixed-effects logit model, N = 316). The coefficients increase in magnitude and become statistically significant with α = .10. Because the model is the same for both sets of results, any differences we see must be due to the change in sample composition. In the next column, we remove the fixed effect. Comparing the results in the second and third columns shows that the coefficient for both measures nearly doubles; both are now statistically significant with α = .05. The only difference in these two sets of results is the model chosen, since the samples are the same. From the logit model to the fixed-effects logit model, the coefficient for “worked 1 to 30 hours” increases by 1.311, and the coefficient for “worked more than 30 hours” increases by 1.008. In this case, we can infer that 52% of the change (0.683 / 1.311) in the “worked 1 to 30 hours” variable is due to sample composition, and 48% is due to the removal of the fixed effect. For the “worked more than 30 hours” variable, 49% is due to sample composition, and the removal of the fixed effect accounts for the remaining change.

In the second set of columns, we report results for a logit model of performance above average. Neither maternal work hours coefficient is statistically significant in the logit model. Next, we report results using a logit model with the subsample that has variation in the outcome. The coefficient for worked more than 30 hours is −0.530 and is statistically significant. Since we have not removed the unobserved time-invariant characteristics in this model, the change in the coefficient is due to change in the sample composition. In the next set of entries, we remove the fixed effect. The coefficient for “worked more than 30 hours” increases to −0.666 and remains statistically significant. From the logit model to the fixed-effects logit model, the coefficient estimates change by 0.588. In this case, 77% of the change observed is due to sample composition, and the remaining 23% of the change is due to the fixed effect.

The result for “parent contacted about behavior problems” is similar, although the removal of the fixed effect is relatively more important to the change in the coefficient estimates from the full sample to the fixed-effects model than in the “performed above average” results. For the “worked more than 30 hours” variable, 56% of the change in the coefficient estimates can be attributed to the sample composition and 44% to the removal of the fixed effect.

Simultaneity Bias

In addition to the issues examined above, we also considered the temporal ordering of the maternal work hours measure and the schooling outcomes. The school participation and performance variables were measured contemporaneously; the reference period for the “number of days late” and “number of days missed” variables is the past four weeks, and for the “skipped school/class” and “contacted about behavior problems” variables, the reference period is the past 12 months. If mothers change their work hours in response to these factors, then our coefficient estimates might suffer from simultaneity bias. In other words, a mother may change her work hours in response to schooling outcomes instead of or in addition to the responses we hypothesize.

Though this is a reasonable concern, if such a bias exists, then our estimates are probably conservative. For example, consider the coefficient estimate for “parent contacted about behavior problems.” We find that increases in maternal work hours are positive and statistically significantly related to this outcome. If being contacted by the school is altering the mother’s work hours, it would probably reduce the hours she works, not increase them. If this countervailing influence exists, then our estimates should be biased toward zero. A similar argument can be made for the “performed above average in school” and “skipped school/class without permission” models.

Income

Before concluding, we address one potential mechanism that has been noticeably absent throughout this analysis: income. We first highlight a couple of points on this issue that are pertinent to our analysis. Several authors who have studied parental income and its influence on children have advocated using the family’s permanent income, which is fixed over time (Mayer 1997; Solon 1992). Since the fixed-effects model removes all time-invariant factors, then the maternal employment estimate should not be biased by the omission of permanent income in the model or by any other factor that is constant over time.

Of course, the maternal employment measure may be capturing changes due to the transitory component of income, which is endogenous to our specification. If maternal employment causes income to change and that income change influences the outcome, then one would not want to include income in the models because it will over-control for the real effects of maternal work hours. If, however, it is income (or lack thereof) that causes mothers to work, and income influences the outcome, then excluding income from our model might bias the estimates of maternal work hours. It is, of course, possible that both influences are important.

In a final supplemental analysis (results available upon request), we reestimated the models presented in Table 3 to include a measure of total family monthly income at the time of the survey (a combination of earned income, nonearned income such as child support, and cash assistance). In general, the coefficient estimates and standard errors for maternal employment are nearly identical to those reported in Table 3, sometimes slightly smaller and sometimes slightly larger, with one exception. The coefficient estimate for working more than 30 hours in the “parent contacted about behavior problems in school” model declines nearly 0.1 and is no longer statistically discernible from zero. These results indicate that for two of the three outcomes, the relationship between maternal employment and the schooling outcomes in this low-income population does not derive primarily from the resulting transitory change in economic resources available to a family when a mother works.

DISCUSSION AND CONCLUSIONS

Using a fixed-effect approach along with controls for a broad array of time-varying characteristics, we find evidence of unfavorable effects of maternal work hours on three of six adolescent school participation and performance outcomes. Compared with the adolescents of mothers who do not work, those of mothers who work are more likely to skip school without permission, whereas only those whose mothers work 31 or more hours are less likely to perform above average and more likely to exhibit behavior problems that result in a call home. In some instances, sons seem to be affected more than daughters by their mothers’ work, with notable increases in incidences of being late for school and decreases in performing above average among boys when mothers work more hours. Importantly, supplemental analyses show that the emergence of statistically significant coefficients in the fixed-effects models results in part from changes in the analytical sample we rely upon for identifying the fixed-effects model (i.e., the sample that has variation over time in our variables of interest) and in part from the reduction of the bias caused by omitted variables. Additionally, supplemental analyses reveal that changes in household income do not account for two of the three maternal employment effects we observe. Both of these caveats should be kept in mind when considering these findings.

The findings from this study generally align with prior research that found neutral to negative effects of welfare and employment policies on adolescent schooling outcomes (Gennetian et al. 2004) and demonstrate that unfavorable effects may exist under considerably different economic and policy conditions than existed prior to the 1996 welfare reforms. The consistency of findings under pre– and post–welfare reform conditions, and across studies with comparably aged adolescents and measured outcomes, is additionally striking in light of differences in the identification strategy in the two studies; Gennetian et al. (2004) relied on pre–welfare reform, experimentally induced changes in employment, while the current study identifies naturally occurring changes produced by some combination of policy, local economic conditions, and time-varying personal and family-level characteristics. This study additionally points to effects on school attendance and schooling-related problems that could not be examined in Gennetian et al. (2004).

The findings in this study appear to differ from those reported in a couple of other recent studies. For example, Chase-Lansdale et al. (2003) found a beneficial effect of maternal transitions into work on adolescent anxiety and psychological distress levels and that transitions off of welfare are associated with increased reading skills among adolescents. Similarly, Ruhm (2006) found small benefits of maternal employment on the cognitive development of economically disadvantaged 10- to 11-year-olds. Such comparisons are, in our estimation, difficult to make because of a lack of comparability of schooling outcomes, differences in the age groups included in each study, and differences in the assessment periods. In addition, the Chase-Lansdale et al. (2003) study focused on maternal transitions with a broad differentiation of work hours, and the study by Ruhm (2006) focused on the cumulative effects of employment over the years of a child’s early development, rather than the more nuanced measure of contemporaneous change in maternal work hours that is the focus of the current investigation. Given these differences, it remains unclear to what extent the findings we present are compatible with or contradictory to the findings reported in these other studies.

Although it is beyond the scope of this paper to isolate the potential range of time-invariant, omitted variables that biased our initial regression results, we do offer a potential explanatory factor in an attempt to direct future inquiries into this topic. As an illustrative example, consider findings from the “contacted about behavior problems” estimation. In the initial model, the coefficient estimate was positive but not statistically significant. In the fixed-effects model, the point estimate for the coefficient more than doubled. One potential omitted variable that would produce such a result is neighborhood context, particularly neighborhood safety. If mothers who work are also living in safer neighborhoods, and neighborhood safety is negatively related to being contacted about problem behavior (perhaps because safer neighborhoods have higher quality schools and fewer negative peer influences), then the omission of neighborhood characteristics could explain the results we see. A similar logic can be applied to the skipping school and school performance outcomes.

As carefully documented in Chase-Lansdale et al. (2002), Gennetian et al. (2004), and Ruhm (2006), the current study also indicates that low-income adolescents whose mothers were on welfare are at high risk of school participation and performance problems, as well as other behavioral and emotional problems that might influence their ability to participate and perform well in school. Together, the findings from all of these studies contribute to a growing body of evidence and knowledge about the effects of maternal work hours on this already high-risk group of adolescents. The findings from the current study suggest that low-income adolescents living in neighborhoods of highly concentrated poverty experience some unfavorable effects from their mother’s employment in the post-1996 welfare policy and economic context. As discussed in greater detail in the appendix, our results may be conservative, since some of the most disadvantaged youths in these families were not included in our analysis sample because they did not meet the eligibility requirements established at the outset of the study. Although we have not isolated the reasons why such changes in employment occurred—whether because of public policy or because of a variety of other personal or labor market factors—and we cannot speak to whether effects occur across other domains of adolescent development, such as their socioemotional and physical health, fertility, and work behavior, the findings from this study point to the importance of considering adolescent development in policy debates about welfare, work, income, and out-of-school care. Studies that seek to better understand how changes in family economic conditions and employment behavior affect family life and children’s development should be appropriately designed to also measure the academic, socioeconomic, and behavioral development of adolescents.

APPENDIX

The analysis sample in this study represents a subsample of those adolescents who were enumerated as being present in the households of women in the 1998 (Wave 1) Urban Change survey. Here, we document the reasons why some children were not included in the analyses. As seen in Appendix Table A1, of the 1,691 Focal B children in the data set, 733 adolescent children were excluded in Wave 1, and 951 adolescent children were excluded from Wave 2. The most common reason for exclusion is that they were older than 18 years by the time of the 2001 (Wave 2) survey (N = 411) and no longer met the inclusion criteria for the study (i.e., the child must have been 12–18 years old in 1998 and no older than 18 years in 2001).

Appendix Table A1.

Criteria to be Eligible for Cross-Sectional Analysis Sample

	Wave 1	Wave 2	Group of Interest
Total Focal B children	1,691	1,691
Children older than 18 at Wave 2	411	411	Age-Ineligible Group
Children whose families were not reinterviewed at Wave 2	286	286	Excluded Group 1
Children whose mothers said they had 0 children living with them	3	32	Excluded Group 2
Children whose mothers said Focal B child was not living with them	14	86	Excluded Group 2
Children who were deceased	5	7	Excluded Group 2
Children not currently in schoola	9	125	Excluded Group 3
Children missing all outcomes for unknown reasons	1
Children missing age in Wave 2	4	4
Number of Children Eligible for the Sample	958	740
Total Child-Year Observations	1,698

^aWhen the mother was asked what the child’s current grade in school was, these children fell into one of the following categories: (1) graduated from high school or earned a GED, (2) in college, (3) not in school, (4) in a GED or an adult basic education program, or (5) refused or responded “I don’t know.” The Focal B schooling outcomes were then skipped for these children. We do not know how many children in excluded Groups 1 and 2 fell into these categories because their mothers did not answer this question about current grade.

Appendix Table A2.

Comparison of Means for Wave 2 Analysis Sample and Children Excluded From the Sample^a

Variable	Comparison of Characteristics Measured at Wave 1		Comparison of Characteristics Measured at Wave 2
	Wave 2 Analysis Sample	Excluded Group 1b	Wave 2 Analysis Sample	Excluded Group 2c	Excluded Group 3d
Child Outcomes
% ever repeated a grade	23.1	22.9	28.0	24.5	38.9*
% ever dropped out of school	3.0	11.9**	2.1	24.4**	46.0**
% ever suspended/expelled from school	30.8	29.3	33.5	42.9†	39.8
% ever had/fathered baby	0.7	5.6**	5.0	16.7**	19.4**
% ever in trouble with police	5.3	9.9*	12.8	26.2**	15.3
Number of days late for school	1.2	0.8**
Number of days missed school	1.7	1.8
% skipped school/class without permission	9.9	15.2*
School performance (scale 1–5)	3.6	3.5
% performed above average in school	52.8	50.0
% performed below average in school	16.7	16.9
% parent contacted about behavior problems in school	42.2	43.7
Characteristics of Children
% male	50.0	51.7	50.0	52.8	47.2
Age	13.6	15.0**	16.4	17.3**	18.1**
Characteristics of Mothers
Age	35.8	36.9**	38.6	38.7	40.9**
% no high school diploma or equivalent	48.1	48.6	43.4	47.2	46.4
% currently married	8.3	11.5	15.0	13.1	16.3
% currently cohabiting	9.6	6.6	12.1	13.7	10.4
CES-D scoree	17.6	17.0	16.2	18.9*	18.3†
% with health limitation	23.1	26.6	24.3	27.6	28.8
SF-12 physical component scoref	46.8	45.9	46.1	45.5	44.4
% physically abused (past year)	7.3	9.7	4.3	8.6	9.5†
% emotionally abused (past year)	39.4	33.3†	25.2	34.2†	42.3**
% used hard drugs (past 30 days)	1.8	2.6	1.9	4.3	1.8
% born in the United States	78.1	75.2	78.1	80.0	77.6
% black	68.8	61.9*	68.8	69.6	64.8
% Hispanic	25.6	30.1	25.6	25.6	26.4
% white	3.9	6.3	3.9	3.2	8.0
% other	1.6	1.7	1.6	1.6	0.8
Total Number of Children	740	286	740	125	125

^aThe sample is children aged 12–18 at Wave 1 and younger than 19 at Wave 2.

^bChildren missing from the sample because of survey nonresponse. Because we do not have any Wave 2 data for these children or their families, comparisons with the analysis sample were made using Wave 1 data.

^cChildren whose families were reinterviewed at Wave 2 but who are missing from the Wave 2 sample because they were not living with their mother at Wave 2 (we do not know if they were in school).

^dChildren whose families were reinterviewed at Wave 2 and who lived with their mother at Wave 2, but who are missing from the Wave 2 sample because they were not in school at Wave 2.

^eCES-D = Center for Epidemiologic Studies-Depression scale; Scores can range in value from 0 to 60. A score above 23 is considered an indicator of a high risk of depression.

^fSF-12 is a health survey with 12 questions designed to measure mental and physical health. Physical component scores have been standardized in a national sample to a mean of 50 and a standard deviation of 10.

^†p < .10;

^*p < .05;

^**p < .01

Of those who were age-eligible in 2001 (N = 1,280), an additional 536 were excluded from the longitudinal analytic sample (the sample used for the fixed-effects models) for one of three reasons. The first reason is survey nonresponse; a total of 286 mothers were not reinterviewed in 2001 (excluded Group 1). The second reason is that the children were not living with their mothers at the time of the 2001 survey (excluded Group 2). A total of 125 children were excluded either because the mother reported that she had no children living with her (N = 32), the mother reported that the age-eligible adolescent child, specifically, was no longer living in her household (N = 86), or the mother reported that the age-eligible adolescent child was deceased (N = 7). An additional 125 children were excluded because the age-eligible child was no longer in school, either because they had graduated by age 18 years or they had dropped out of school (excluded Group 3). Finally, four adolescents were excluded because we could not determine their ages at the Wave 2 survey, and we were unsure whether they met the sample selection criteria.

Although most of the longitudinal school participation and performance outcomes are missing for the groups of adolescents who were excluded in the study, we were able to compare these adolescents to the adolescents in the analysis sample by using other survey outcomes and characteristics of the children and their mothers. Appendix Table A2 shows that all of the groups of children excluded from the Wave 2 survey sample were more likely to have dropped out of school and to have had or fathered a baby.11 In addition, excluded Groups 1 and 2 were more likely to have ever been in trouble with the police than the children in the analysis sample at Wave 2.12 Children excluded because they were no longer living with their mothers or because they were no longer in school also had mothers who were more likely to be depressed and to be emotionally, psychologically, or physically abused.

These results provide some evidence that children in the worst circumstances and with the poorest outcomes were ineligible for inclusion in the analysis sample. This suggests that the analysis sample represents a slightly better-off or less at-risk sample of adolescents and that our findings of unfavorable effects of maternal employment on adolescent school participation and performance over time might be conservative estimates.